by Hao AI Lab
Open-weight 4-step sparse-distilled version of MiniMax H3 for fast text-to-video-and-audio generation on Blackwell GPUs.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
FastH3 Preview v1 is a post-trained, distilled variant of MiniMax's open-weight H3-Base video model, released by Hao AI Lab's FastVideo project. It targets text-to-video-and-audio (T2VA) generation and was developed with Nuva Lab, NVIDIA FastGen and NVIDIA Enterprise Products teams. It distills H3's base transformer to 4 steps and reuses the H3-Base text encoder, video and audio VAEs, tokenizers and schedulers.
The recommended checkpoint is a 4-step VSA / Data-Free model: it was trained from prompts only (no target videos) using Video Sparse Attention (90% sparse, tile 64). Per the authors, it can generate a 15-second 768p clip in under 13 seconds on 8×B200 GPUs, with up to 14x speedup on a single Blackwell GPU. It supports variable resolution, aspect ratio and duration in one checkpoint (heights and widths in multiples of 32), and outputs stereo audio with video.
Full weights and a pre-extracted LoRA are provided, along with dense and synthetic-data ablation checkpoints. The recommended checkpoint and LoRA require FastVideo's VSA-H3 attention backend and kernel; dense attention is not a drop-in substitute. Preview v1 supports T2VA only; first/last-frame (FL2VA) and reference (Ref2VA) checkpoints are in development, and training code is listed as coming soon. A community ComfyUI checkpoint (about 22.9 GB) was reported by a YouTube creator.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Speedup of MiniMax H3 using DMD2 distillation and video sparse attention, up to 14x faster. On 8 B200s it generates a 15-second clip in under 30 seconds; Kijai released a 22.9GB ComfyUI checkpoint.
Closest neighbours by meaning, not just by category.
Model · Hao AI Lab
Distilled MiniMax H3 reference-to-video-and-audio model from Hao AI Lab, generating in 8 steps instead of 50.
Model · MiniMax
Open-weight video generation model from MiniMax, part of its H3 family alongside H3 Image.
Model · MiniMax
Open-weights MiniMax image generation and editing model producing up to ~16 MP images in 6 sampling steps.
Model · Unsloth
GGUF quantizations of MiniMax-H3 video-with-audio generation model, for stable-diffusion.cpp and Unsloth.
Research · MiniMax
Open-weight multimodal model from MiniMax whose video pipeline can be run audio-only to generate music and sound effects.
Model · Kijai
Community-quantized W4A8 and int8 VAE checkpoints of MiniMax H3 video model for ComfyUI, by Kijai.